It's essentially by design. To keep AI from destabilizing, the way AI is tuned is to essentially 'go the same safe way' on similar prompts. LLM models are essentially a giant field of weights for chains of tokens. Prompts are turned into tokens and the closest result is the output. The more room you give it to vary, the more different styles and variation it can express with a propmpt, but it raises the chance of destabilizing the ouput, producing things like additional fingers or complete garbage. It's tuned to walk the same way to produce something we perceive as "passable" but it comes at the cost of taking away variation, which is why AI art/writing will consistenly converge towards the same narrow style set.

(This explanation is not super accurate. I just tried to convey the jist of it)